# Plan Agent Capacity And Cost > Work out whether an agent's message capacity, quota or spend is sized correctly, before a limit gets hit or a bill surprises someone. Use when a Copilot Studio agent is close to its message capacity or credit allocation, when Foundry model quota or token spend needs to be forecast for a launch, when a finance stakeholder asks what an agent costs and nobody can answer in real numbers, when pay-as-you-go and prepaid credits need to be reconciled, or when usage is growing and nobody has planned for what happens at ten times the current volume. - Skill: `ragnarpitla/plan-agent-capacity-and-cost` (Agent Skill) - Install (CLI): `npx skillmds@latest add ragnarpitla/plan-agent-capacity-and-cost` - Raw SKILL.md: https://api.skillmd.com/api/skills/ragnarpitla/plan-agent-capacity-and-cost/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Product & Planning - Author: RagnarPitla (https://skillmd.com/u/ragnarpitla) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/ragnarpitla/plan-agent-capacity-and-cost --- Capacity and cost problems are almost always predictable in hindsight and invisible in advance, because nobody was watching the number until it became a problem. This skill is about watching it on purpose: sizing capacity before launch, and knowing what a given level of usage actually costs before finance asks. This is a budgeting and forecasting concern, not a health signal - if the question is "is the agent behaving normally", that is `monitor-agent-telemetry`. If the question is "can we afford what it is about to do, or what happens when a limit is hit", that is here. ## Copilot Studio: credits, messages and capacity Copilot Studio usage is metered in Copilot Credits, consumed differently by message type and feature (a classic answer, a generative answer, an agent action, tool or voice usage each consume a different amount). Credits come from either a prepaid capacity pool shared across the tenant, or pay-as-you-go billed to an Azure subscription once prepaid capacity is exhausted. Practical questions worth asking before capacity becomes an incident: - Is consumption tracked per environment and per agent in the Power Platform admin center, so a runaway agent is visible before the whole tenant's pool is affected? - Is the chatbot only activated on user interaction (a click), rather than on page load? Loading a chat widget can itself generate billable messages at a scale nobody intended. - Is there a plan for what happens automatically when prepaid credits run out - silent failure, a hard stop, or an agreed switch to pay-as-you-go? That should be a decision made in advance, not discovered mid-incident. ## Foundry: quota and token spend Foundry model quota is managed at the subscription level (pooled across regions for a given model and version, either globally or by data zone), so forecasting needs to account for everything sharing that pool, not just one agent's expected traffic. Ask early, before a launch: - What is the expected token volume per conversation, and does the requested quota actually cover peak concurrent usage, not just average? - Would provisioned throughput (a reserved capacity unit, billed for what is reserved rather than per token) be cheaper than pay-per-token at this agent's expected sustained volume? This is usually a real trade-off, not a default choice either way. - Is there a budget and an alert configured before launch, not added reactively after the first surprising invoice? ## Forecasting growth, not just today The number that matters is not current usage, it is current usage at the volume expected after the next rollout, integration or marketing push. Ask directly: if usage grew ten times, does the current capacity plan survive that, or does someone need to act before it happens rather than during a degraded experience? ## Reconciling the bill with what it was for When a stakeholder asks "what does this agent cost", the honest answer usually needs both the platform's own usage view (Power Platform admin center for Copilot Studio, the Foundry or Azure portal for Foundry) and cost management tooling, because credits, provisioned throughput and pay-as-you-go usage are not all visible in the same place. Reconcile before answering rather than quoting whichever number was easiest to find. ## Do not - Recommend buying more capacity as the first response to hitting a limit, without first checking whether the consumption itself is expected (versus a widget firing on page load, a retry loop, or a runaway automation). - Treat provisioned throughput as automatically cheaper. It is a trade of guaranteed capacity for guaranteed spend, and it only wins at genuinely sustained volume - check the actual usage pattern before recommending it. - Answer a cost question from a single dashboard without checking whether it covers pay-as-you-go, prepaid and provisioned spend together. A partial number stated confidently is worse than saying it needs reconciling. - Set a budget alert as an afterthought after a launch. It belongs in the same conversation as the capacity plan, before traffic arrives. ## Sources Verified as resolving on 2026-08-19. Review by 2026-11-17. - Manage Copilot Studio credits and capacity: https://learn.microsoft.com/en-us/power-platform/admin/manage-copilot-studio-copilot-credits-capacity - Microsoft Copilot Studio billing rates and management: https://learn.microsoft.com/en-us/microsoft-copilot-studio/requirements-messages-management - Microsoft Foundry Models quotas and limits: https://learn.microsoft.com/en-us/azure/foundry/foundry-models/quotas-limits - Plan and manage costs for Microsoft Foundry: https://learn.microsoft.com/en-us/azure/foundry/concepts/manage-costs Credit rates, quota scoping and pricing pages change more often than most Microsoft Learn content, hence the shorter review window above rather than the default. If a link 404s, the page was probably renamed - say you could not verify it rather than guessing a replacement URL, and treat any rate quoted here as illustrative rather than current until you have re-checked it.